GPT-5.6 Sol

编程9 项

• ProgramBenchVals AI1.5%+
• CyberBench Patch v1.1Vals AI89.3%+
• Terminal-Bench 4.0Artificial Analysis39.9%+
• Terminal-Bench 4.0Vals AI37.9%+
• Code MigrationVals AI52.9%+
• LiveBench CodingLiveBench83.9%+
• Vibe Code Bench v1.1Vals AI80.5%+
• Vibe Code Bench 1–100Vals AI20.0%+
• SciCodeArtificial Analysis57.1%+

科研9 项

• FrontierMath T1–3Epoch AI89.1%+
• Mystery Game PuzzlesEpoch AI58.0%+
• Chess PuzzlesEpoch AI59.0%+
• CritPtArtificial Analysis32.3%+
• Terminal-Bench ScienceVals AI20.0%+
• LiveBench ReasoningLiveBench91.7%+
• Terminal-Bench ScienceArtificial Analysis22.4%+
• MysteryMechanismVals AI33.3%+
• Humanity's Last ExamArtificial Analysis49.5%+

专业办公8 项

• AA-Analyst AgentArtificial Analysis47.5%+
• LiveBench Data AnalysisLiveBench79.8%+
• AA-BriefcaseArtificial Analysis1,478.5+
• AutomationBench-AAArtificial Analysis60.1%+
• GDPval-AAArtificial Analysis1,607.2+
• Finance AgentVals AI53.8%+
• Harvey 法律 Agent 基准Vals AI2.5%+
• MedCodeVals AI44.0%+

知识问答7 项

• SimpleQA VerifiedEpoch AI69.7%+
• MLCR-AAArtificial Analysis30.0%+
• LiveBench LanguageLiveBench87.7%+
• AA-OmniscienceArtificial Analysis22+
• LiveBench IFLiveBench71.8%+
• AA-LCRArtificial Analysis84.0%+
• PBullshitBench v2BullshitBench47.0%+

真人盲选1 项

• Arena TextLMArena1,485.1+

测过，但按规则不计入

这些公开成绩未满足本榜的可比条件，具体原因如下。统一固定工具、全程使用同一型号可以采用；多模型回退、混合模型与专属异构 Agent 系统不混入模型榜。资料未说明的项目保留未知。

• APEX-Agents 1.1：Pro 系统多次采样后汇总作答，不是 GPT-5.6 Sol 本身。
• τ³-Banking · Mercor：Pro 系统多次采样后汇总作答，不是 GPT-5.6 Sol 本身。